From 4bec6f440754ce8037d249c38f2316be0c266f06 Mon Sep 17 00:00:00 2001 From: gasoonjia Date: Sat, 24 Oct 2026 00:13:44 -0700 Subject: [PATCH] Fetch Muse Glimmer serving tokenizer files from the non-GGUF repo The CUDA solo-text-batching serving smoke downloaded tokenizer_config.json and chat_template.jinja from ${HF_MODEL} (meta-models/Muse-Glimmer-30B-GGUF). That repo only ships GGUF weights, so the download 404s and the job has failed on every main run since #24010, right after the batching correctness check passes. Fetch both files from $TOKENIZER_URL (meta-models/Muse-Glimmer-30B), which is where tokenizer.json already comes from. Use `curl -f` so a missing file still fails the job. Test plan: ran the new download loop locally against $TOKENIZER_URL; both files downloaded and tokenizer_config.json parses as JSON. `bash -n` passes. CI: test-muse-glimmer-cuda-e2e-kquant-17gb-solo-text-batching (ciflow/cuda). --- .ci/scripts/test_model_e2e.sh | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/.ci/scripts/test_model_e2e.sh b/.ci/scripts/test_model_e2e.sh index 6d63a37b728..786cd81f9e2 100755 --- a/.ci/scripts/test_model_e2e.sh +++ b/.ci/scripts/test_model_e2e.sh @@ -618,7 +618,9 @@ if [ "$DEVICE" = "cuda" ] && [ "$MODEL_NAME" = "muse_glimmer" ] && [ "$MODE" = " --mode requests --requests 5 --prompt_tokens 600 --max_new_tokens 64 --stagger_ms 200 pip install -r examples/llm_server/python/requirements.txt - python -c "from huggingface_hub import hf_hub_download; [hf_hub_download('${HF_MODEL}', f, local_dir='${MODEL_DIR}') for f in ('tokenizer_config.json', 'chat_template.jinja')]" + for f in tokenizer_config.json chat_template.jinja; do + curl -fL --retry 3 --retry-all-errors "$TOKENIZER_URL/$f" -o "$MODEL_DIR/$f" + done PORT=$(python - <<'PY' import socket